Uav Flight Data Instantaneous Anomalies Detection Based On Subspace Learning

The accurate and timely perception of the Unmanned Aerial Vehicle(UAV)’s own health status is the premise of safe and reliable flight.It is also the basis for achieving complete autonomous flight in the future.Flight data are series of collected parameters reflecting the real flight status of UAV.The anomalies in flight data usually mean abnormal flight status of UAVs.Flight data anomaly detection technology helps the state monitoring of drones and the mining of potential anomalies.Therefore,the research on the flight data anomaly detection has received extensive attention in recent years and has become one of the key technologies for onboard system health management.However,the classic flight data anomaly detection methods still face great challenges when it comes to the onboard UAV health management requirements.First of all,most of the current researches focus on the mining of offline flight databases.However,in order to meet the health monitoring needs of the UAV,it is necessary to explore online instantaneous anomaly detection methods.Secondly,the labled data samples are not easy to obtain.How to use the unsupervised machine learning methods to fully exploit the structure of the data itself is still very difficult.In addition,sufficient decision information is needed for emergency response measures such as flight control rate reconstruction or task reconstruction.The flight data anomaly detection method not only needs to detect the abnormal state at the current time,but also to provide explanatory information such as the abnormal source.Aiming to solve above problems,this paper proposes an unsupervised subspace learning anomaly detection framework with low rank and sparse constraints,and conducts an in-depth study on online representation of multidimensional flight data,online detection of outlier point and online detection of anomaly sources in the task of flight data instantaneous anomaly detection.The main research and contributions are as follows:(1)The traditional flight data representation methods do not fully exploit the structural characteristics of the data itself.Subspace learning also faces the challenge of local time series information learning.In this work,the flight data representation and the anomaly detection task are integrated in a learning framework,and a method based on onlinesubspace tracking(Flight data Representation on Online Subspace Tracking,FROST)is proposed.First,the data subspace matrix is used as a low dimensional representation of the raw flight data.It is also learned in an online unsupervised manner with the criterion of minimizing the reconstruction error.Secondly,the learned subspace preserves more local time series information of the original data,which makes it has better data reconfigurability and abnormal sensitivity.For normal data,the estimated subspace matrix reconstructs the original data with a lower error;for the anomalous data,the estimated subspace matrix distribution changes,thereby facilitating the discrimination of the abnormal samples.In addition,the data subspace is calculated in a recursive manner with a small time and space complexity,which is suitable for online processing of flight data streams in onboard scenario.The experimental results show that the proposed method is superior to the similar methods for flight data and has good data representation ability and abnormal sensitivity.(2)For the traditional flight data instantaneous anomaly detection methods,there are limitations such as the dependence on prior distribution assumption,the difficulty to determine the neighborhood,and the problem of lightweight representation and multidimensional spatiotemporal relationship modeling.An unsupervised flight data outlier dectection method is proposed,which is based on partial subspace learning algorithm(Partial Subspace Learning Algorithm for Online Anomaly Detection,PSLOAD).First,PSLOAD models the spatiotemporal relationship between multidimensional flight data using partial subspaces constrained by low rank.Secondly,by using the similarity characteristics of the data subspace in the timing direction,PSLOAD improves the solution of the subspace vector in the manner of projection approximation,and obtains the representation with abnormal discriminability.Finally,the degree of anomaly is determined by measuring the subspace vector variation of the target instance.Compared with the classical flight data outlier detection model,PSLOAD does not require a priori distribution assumption and predefined data sliding window.It is with the characteristics of high accuracy,lower false detection rate and lower run time overhead.(3)In order to solve the problem of the missing data source information in the traditional flight data instantaneous anomaly detection methods,this work proposes a Structured Sparse Subspace Learning Anomaly Detection Algorithm(SSSLAD)algo-rithm.First,SSSLAD uses a spatial dependency between different flight data to design a predefined structured sparse norm to preserve data source information.Then the anomalous source detection is quivalents to optimize the structured sparse subspace.Secondly,in order to solve the problem of structural sparse subspace learning,the original nonconvex nonsmooth optimization problem is transformed into a stepwise smooth convex problem.The predefined structured sparse norm will induce the subspace projection coefficient matrix to belong to the pre-specified sparse mode,which improves the mixed nature of the subspace,thus achieving the correct detection of the anomaly source.In addition,the Nesterov momentum optimization method,the Euclidean projection method and the similarity characteristics of the subspaces in adjacent time intervals are fully utilized to accelerate the convergence of the structured sparse subspace learning problem.The experimental results show that compared with the classical anomaly source detection model,SSSLAD has higher accuracy and lower false detection rate,and reduces computation time.